跳转至

文章背景与核心概要

传统的测试时扩展方法（例如基于置信度的投票）主要依赖于内部信号（如词元对数概率）来对并行的大语言模型（LLM）输出进行加权。然而，当这些方法应用于需要检索并依赖外部文档的现代多轮搜索智能体时，置信度得分会因为复制膨胀（copy inflation）现象而失效——即从检索到的上下文中复制的词元会获得人工放大的对数概率。

为了解决这一问题，作者提出了检索扎根投票（Retrieval-Grounded Voting, RGV）。该方法完全绕过了词元概率，转而通过计算最终答案与检索到文档之间的词汇重叠度来对输出进行评分。在不需要额外 LLM 调用的前提下，RGV 在多个基准测试和模型上实现了稳定的性能提升，为多轮搜索智能体的可靠性扩展提供了一条新途径。

Beyond Confidence: Test-Time Scaling for Multi-Trust Search Agents via Retrieval Grounding

Summary

Traditional test-time scaling methods (like confidence-based voting) rely on internal signals such as token log probabilities to weigh parallel LLM rollouts. However, when applied to modern multi-turn search agents that retrieve and condition on external documents, these confidence scores break down due to copy inflation—where tokens copied from retrieved context receive artificially inflated log probabilities. To solve this, the authors propose Retrieval-Grounded Voting (RGV), which bypasses token probabilities entirely by scoring rollouts based on the lexical overlap between the final answer and the retrieved documents. RGV achieves consistent performance gains across multiple benchmarks and models without requiring additional LLM calls.

元数据

Metadata

  • 学科分类: 人工智能 (cs.AI)
  • Subjects: Artificial Intelligence (cs.AI)
  • 录用情况: EMNLP 2026 Findings 录用
  • Accepted to: EMNLP 2026 Findings
  • 作者团队:
  • Hyunho Kook
  • Junhyuk So
  • Tianyu Fu
  • Haizhong Zheng
  • Beidi Chen
  • Authors:
  • Hyunho Kook
  • Junhyuk So
  • Tianyu Fu
  • Haizhong Zheng
  • Beidi Chen
  • 提交日期:
  • 2026年8月25日提交 (v1)
  • 2026年8月27日最后修订 (v2，当前版本)
  • Submission Dates:
  • Submitted on 25 Aug 2026 (v1)
  • Last revised 27 Aug 2026 (v2, this version)

摘要

Abstract

基于置信度的投票通过使用诸如词元对数概率等内部信号对每个并行 LLM 输出进行加权来聚合结果，这在单轮推理中已被广泛研究。然而，现代 LLM 越来越倾向于作为多轮搜索智能体，执行检索外部文档并基于其进行条件生成。

Confidence-based voting aggregates parallel LLM rollouts by weighting each with internal signals such as token log probabilities, and has been actively studied for single-turn reasoning. However, modern LLMs increasingly act as multi-turn search agents that retrieve and condition on external documents.

在本文中，我们发现基于置信度的投票很难很好地迁移到这种多轮设置中，并识别出其背后的失效原因：复制膨胀。当检索到的文档被追加到智能体的上下文中时，从这些文档中复制的词元会系统性地获得放大的对数概率。这会拉平每个问题内的置信度得分，从而削弱最终加权投票的效果。

In this paper, we show that confidence-based voting transfers poorly to this multi-turn setting, and identify the underlying failure reason as copy inflation: when retrieved documents are appended to an agent's context, tokens copied from those documents receive systematically inflated log probabilities. This flattens confidence scores within each question and weakens the resulting weighted vote.

为了解决这个问题，我们提出了检索扎根投票（RGV）。它通过计算最终答案与智能体检索到的文档之间的词汇重叠度来对每个输出进行评分。通过在受污染上下文之外计算信号，RGV 同时避开了词元对数概率和额外的 LLM 调用。在四个搜索智能体基准测试和五个 LLM 上，RGV 的表现始终优于基于置信度的投票，准确率最高提升了 +5.4%，在少数正确问题（即正确答案仅出现在 8 个输出中的 1-2 个里的问题）上的准确率提升高达 +35%。

To address this issue, we propose Retrieval-Grounded Voting (RGV), which scores each rollout by the lexical overlap between its final answer and the documents it retrieved. By computing the signal outside the contaminated context, RGV sidesteps both token log probabilities and additional LLM calls. Across four search-agent benchmarks and five LLMs, RGV consistently outperforms confidence-based voting, with gains of up to +5.4% accuracy and +35% on minority-correct questions (where the correct answer appears in only 1-2 of 8 rollouts).

链接与资源

license icon